CulturalBench: a Robust, Diverse and Challenging Benchmark on Measuring the (Lack of) Cultural Knowledge of LLMs

2024-10-03

arXiv.org

AI Summary

Existing cultural knowledge evaluation benchmarks for language models suffer from uneven geographical coverage, monolithic question design, and inadequate modeling of multi-answer questions. Method: We introduce CulturalBench—the first global, multicultural evaluation benchmark—comprising 1,227 manually authored, five-stage-verified questions spanning 45 regions (including underrepresented ones like Bangladesh and Zimbabwe) and 17 cultural domains. It employs a dual-difficulty (Easy/Hard) evaluation paradigm to systematically probe model sensitivity to question phrasing and convergence bias under multiple valid answers. Contribution/Results: CulturalBench reveals pronounced LLM deficiencies in South American and Middle Eastern cultural knowledge. Experiments show GPT-4o achieves 61.5% accuracy on the Hard subset (vs. human 92.6%), while Llama3-8b scores only 21.4%. OpenAI models outperform others across all regions except Oceania.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Question Answering
• Search and Optimization: Evaluation and Analysis

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

To make large language models (LLMs) more helpful across diverse cultures, it is essential to have effective cultural knowledge benchmarks to measure and track our progress. Effective benchmarks need to be robust, diverse, and challenging. We introduce CulturalBench: a set of 1,227 human-written and human-verified questions for effectively assessing LLMs' cultural knowledge, covering 45 global regions including the underrepresented ones like Bangladesh, Zimbabwe, and Peru. Questions - each verified by five independent annotators - span 17 diverse topics ranging from food preferences to greeting etiquettes. We evaluate models on two setups: CulturalBench-Easy and CulturalBench-Hard which share the same questions but asked differently. We find that LLMs are sensitive to such difference in setups (e.g., GPT-4o with 27.3% difference). Compared to human performance (92.6% accuracy), CulturalBench-Hard is more challenging for frontier LLMs with the best performing model (GPT-4o) at only 61.5% and the worst (Llama3-8b) at 21.4%. Moreover, we find that LLMs often struggle with tricky questions that have multiple correct answers (e.g., What utensils do the Chinese usually use?), revealing a tendency to converge to a single answer. Our results also indicate that OpenAI GPT-4o substantially outperform other proprietary and open source models in questions related to all but one region (Oceania). Nonetheless, all models consistently underperform on questions related to South America and the Middle East.

Problem

• Assessing cultural knowledge in language models globally
• Evaluating LM performance on underrepresented cultural regions
• Identifying LM biases in handling multiple correct answers

Innovation

• Human-AI teaming creates diverse cultural questions
• Five annotators verify each question for accuracy
• Benchmark challenges LMs with tricky multiple-answer questions